# CAIN-2027 Evolution 103: cross-domain transfer and skills, measured

Preregistered (`PREREGISTRATION.json`) before the sealed runs. Gate counts: {'PASS': 2, 'PARTIAL': 3, 'FAIL': 3}. Not AGI or VAGI. No external reviewer.

## What was built
- A second vertical, **service configuration debugging** (`cain/cain2027/e103/service.py`): two knobs, a hidden target configuration reported only through a latency-SLO signal, scrambled control labels (p 0.8), propagation delays (p 0.6) and structured incompatible configurations. It shares the action and observation protocol with the grid worlds, so the transfer measured is **near transfer** between structurally analogous domains, not between modalities.
- A **skill library** (`skills.py`): opening procedures mined from verified experience, each with a precondition, an internal postcondition, provenance, validation evidence, a risk class and an expiry rule; counterexample-guided truncation; acceptance only when a skill beats no-skill on validation. Skills carry no authority: every skill action passes the governance gate.

## Results
| test | result |
|---|---|
| Second vertical (service configuration) | **PASS**: built and solvable: in-domain organism vs no memory reference 130 vs 136 of 200 solved; mean steps 38.945 vs 39.595; faster on 82, slower on 46 (p = 0.0019) |
| Cross-domain transfer grid -> service (T1) | **PARTIAL**: 133 vs 136 of 200 solved; mean steps 39.545 vs 39.595; faster on 78, slower on 52 (p = 0.028); classified POSITIVE by the preregistered sign test, but no success gain and a negligible mean-step change |
| Cross-domain transfer service -> grid (T3) | **PASS**: 80 vs 80 of 80 solved; mean steps 14.113 vs 15.812; faster on 56, slower on 14 (p = 4.3e-07); POSITIVE |
| Negative-transfer guard (T2, invariance vs basic) | **PARTIAL**: 133 vs 125 of 200 solved; mean steps 39.545 vs 40.62; faster on 60, slower on 53 (p = 0.57); not significant |
| Skill acquisition (validation) | **PARTIAL**: accepted identify [0, 1] and localize [0, 1, 3] on validation; 6 candidates rejected; refinement truncated failing candidates |
| Skills on sealed held-out tasks (S1) | **FAIL**: 130 vs 132 of 160 solved; mean steps 25.869 vs 25.719; faster on 27, slower on 29 (p = 0.89) |
| Skill composition (S2) | **FAIL**: 38 vs 40 of 40 solved; mean steps 21.35 vs 16.3; faster on 17, slower on 22 (p = 0.52); composed is numerically slower |
| Cross-domain skill transfer (S3) | **FAIL**: 72 vs 72 of 100 solved; mean steps 38.18 vs 37.08; faster on 35, slower on 40 (p = 0.64); NONE |

## Reading
- **Memory transfers between the two domains, modestly.** Service experience made the grid hidden-goal tasks clearly faster; grid experience made service incidents faster on more tasks than slower, but with no success gain and almost no change in mean steps.
- **Skills did not survive the sealed test.** The validation rule accepted any mean improvement, and many candidates were tried, so the accepted skills were partly selected by chance. On fresh tasks they gave no gain, and the composed pair was slower on the held-out composition. This is a FAIL, and the acceptance rule needs a significance requirement and a separate confirmation set (E104).
- Grid memory reused on grid hidden-goal tasks (families 2 and 9) is in-domain and is not counted as transfer.

## Verify
`python3 -I verify_e103_offline.py .` in the bundle recomputes every comparison from `rows.json` (stdlib only).

Team: Vincent Noriega (founder), Lena Noriega, Monica Diaz.
